Pinned Tweet
Major ATH tokenomics update: we’re introducing daily ATH burns, funded by a share of network fees.
Every burn goes to a public burn address, where tokens are permanently removed from supply. The mechanism goes live in November.
We’re also replacing the fixed platform fee with a floating network fee, set by a published formula that reads market conditions. When the market weakens, the fee falls and the share that gets burned rises.
Cloud Hosts are protected by design: the fee can never exceed the value of Cloud Host rewards.
The underlying host emissions schedule remains unchanged. The update changes the fee mechanism applied around those rewards.
A dashboard tracking daily, monthly and cumulative ATH burned will be available in November.
Burns that follow real network activity, and you can check every one of them.
Read the full announcement below👇
When a firm that tracks chips for a living sees 45% of its own AI bill go to persistent agents since February, that's a demand signal. Agents keep models loaded and tokens flowing around the clock. That means always-on GPU capacity, which is exactly what Aethir's distributed GPU cloud is built to serve.
Persistent agents went from 0% to 45% of SemiAnalysis's AI spend since February.
That's a firm that tracks chips and data centres for a living, and this chart is its own bill.
The rest is still Claude, OpenAI and xAI used the normal way. The agent share was a sliver through the summer and jumped in the last few weeks.
A persistent agent doesn't wait for someone to type a prompt. It works on a task for hours, splits it into sub-agents, opens a real browser and file system, and runs inference the whole time.
So a team's AI bill starts to follow how much work it hands off, around the clock.
That kind of demand compounds, and it compounds faster the more hours of the day the agents are left running
AI agent graphs turn one request into dozens of model calls. Every node that calls a model is a GPU job.
Anthropic measured ~4x the tokens of a chat for a single agent and ~15x for a multi-agent system.
Gartner expects agentic inference costs per workflow to rise 5x+ through 2028.
Parallel branches help with speed. UC Berkeley LLMCompiler research cut latency up to 3.7x vs a step-by-step agent.
Aethir runs that inference across 430K+ GPU containers in 94 countries, with Aethir Mesh serving open-weight models through one API.
Explore our new blog to learn more👇
Aethir retweeted
Hyperscalers are stretching how long their servers stay in service.
Meta moved to useful lives of about 5.5 years, which keeps older accelerators on the books, and in the fleet, deep into later GPU generations.
That helps an accounting line. It does less for a buyer who needs current inference performance today.
Aethir ACCELERATE plans to build new capacity around NVIDIA B300 and GB300 clusters, sized at 64 to 256 nodes for the workloads enterprises are running right now, and designed to go from site selection to live capacity in months.
Capacity coming online in 2026 should run on 2026 silicon.
Every AI data center announcement is quoted in megawatts.
That is an input number. Two halls drawing the same power can differ 10x in tokens delivered.
Here is the metric replacing it, and what it hides🧵
Power Usage Effectiveness (PUE) determines how much of a grid connection becomes compute.
The industry average is 1.52, flat for seven years. At that ratio, 100 MW of IT load is a 152 MW ask.
NVIDIA quotes 1.1 for a Vera Rubin hall. The 42 MW gap is capacity.
1
8
733
Three questions for any provider: What interactivity level, where is the meter, and what utilization does the number represent?
Full breakdown: aethir.com/blog-posts/aethir…
2
664
Anthropic’s amended S-1 puts a number on the overflow premium: $5B of short-term compute in Q3 2026, bought above multi-year rates.
Here’s what AI labs pay when demand outruns the plan🧵
TechCrunch names neocloud deals with SpaceX, Nscale, Akamai and Volta.
SpaceX alone runs $1.25B a month.
Flexible, ready capacity carries a compute premium.
1
7
1,073
On-demand GPU capacity that’s already running is the cheaper overflow.
Aethir: 430K+ GPU containers, 94 countries, 48-hour provisioning for selected GPU types.
Aethir ACCELERATE plans to build mid-sized data center capacity against signed demand.
Full breakdown: aethir.com/blog-posts/the-ov…
9
1,080
Customers prepaying more than the full capex is about as clear a demand signal as you get. Ready GPU capacity is the scarce thing right now, and whoever brings it online fastest wins. That's the gap Aethir is built to fill.
Nebius customers are now paying more than the full cost of the build up front.
On deals signed in Q3, prepayments covered over 100% of the capex behind them. In Q2 it was 50% to 60%.
More than 90% of the contract value signed in the quarter came with a prepayment, and customers have committed over $12B since the start of Q3.
A lot of people have been asking how the buildout gets paid for. This is a big part of the answer. When capacity is this tight, the customer pays for the GPUs, the data centre and the networking before it starts using them.
For 2027 they expect prepayments and financing against existing investment grade contracts to cover 75% to 85% of $40B to $45B of capex.
In my opinion this is what a supply constrained market looks like from the inside
The Aethir tokenomics update is live.
The ATH burn dashboard is now available to everyone, along with the data on the daily ATH burns.
Here’s how it works🧵
Every completed burn is publicly verifiable.
Dashboard: dashboard.aethir.com/protoco…
Burn contract: arbiscan.io/address/0xf8ec8e…









